"Why Coase needs Hayek" is an essay by Rohit Krishnan, an independent researcher and co-author of the MarketBench paper, published on 2026-05-02 on the Strange Loop Canon Substack (roughly 3,500 words). It extends MarketBench's empirical work into a head-to-head comparison of three ways to coordinate multiple AI models on a task, run across 15 hand-written tasks, and argues that a market mechanism beats a hierarchical orchestrator on cost while matching a single strong model on quality.
Argument
The essay contrasts three topologies for coordinating multiple AI models on a task:
- Solo — one strong model handles everything; the hard part is coherence under one global state.
- Hub-spoke — a hub model decomposes the task, delegates to workers, red-teams, and revises; the hard part is decomposition, since the hub must know what subtasks exist and which model can solve each.
- Market — every available model bids for each task, one wins, the answer is judged, and reputations update across the run; the hard part is allocation and retry, since agents must know how much to bid and on what.
Krishnan frames the choice through Coasean and Hayekian economics. Coasean firm theory predicts that as transaction costs decline, work unbundles into many small firms that must coordinate; how they coordinate — central planning versus price signals — is a Hayekian question. In Krishnan's experiment the price signal wins, which he summarizes as "Coase needs Hayek."
Krishnan ran all three topologies across 15 hand-written tasks (5 coding, 5 reasoning, 5 synthesis). Aggregated across tasks, hub-spoke spent $5.33 averaging 6.7/10, solo spent $1.69 averaging 7.2/10, and the market spent $1.34 averaging 7.2/10, leaving the market roughly 4× cheaper than hub-spoke and matched with solo on quality.
Per-domain findings
On the 5 coding tasks, solo wins on most. Krishnan attributes this to coding rewarding holding the whole class, edge cases, and invariants in one place. Hub-spoke wins only when decomposition is naturally clean, as in coding-003, a refactor task in which cleaning validation, discount logic, and result assembly each go to a separate worker. The market hurts itself by routing 9 of 15 coding runs to GPT-5.2, and 4 runs never fill at all because no bidder takes the job. Krishnan summarizes the pattern as "models are better coders than they are good TPMs."
On the 5 reasoning tasks, the market wins by a wide margin, scoring 7.1 against solo's 5.1 and hub-spoke's 5.2. On a brittle exact-match-probability problem, the correct answer (10/33) appeared only in market runs. Krishnan attributes this to brittle reasoning rewarding independent retries with diverse failure modes, so that one bad first answer does not end the run because another worker can take a shot.
The 5 synthesis tasks are described as the "ambiguous middle." The market shows benefits over hub-spoke, with framing and tradeoff care mattering more than statefulness or pure retry diversity.
Why hub-spoke loses
Krishnan argues that hub-spoke is not simply an org chart with autonomous workers: the hub must solve two problems before workers can solve anything — what the subtasks are and what good recomposition looks like. If either step is wrong, individually competent workers still produce a worse final answer. Hub-spoke wins only when decomposition is obvious enough that the orchestrator does not burn its advantage figuring it out. This, in his account, is why hub-spoke costs roughly 4× as much as the market: the hub does extra reasoning work that does not always add value, and the spokes still spend tokens.
Why markets win despite bidder incompetence
Krishnan describes the market in the experiment as "still the underpowered version of itself, a bartering shantytown rather than the modern New York City," because agents are, in his words, catastrophically bad at knowing how to bid. He cites MarketBench for the finding that Gemini is overconfident, GPT underconfident, and none predicts what it would take to solve a problem. He argues the market wins despite this through two mechanisms: a diversity premium, in which different models — even when trained similarly — diverge enough to give different solutions, exposing the principal to that diversity; and failure detection plus retry, in which the bidding-and-retry loop converts a failed answer into another shot.
Krishnan suggests Recursive Language Models (RLMs, arXiv:2512.24601), which let a model search and update its context, should make markets perform even better, "precisely because of the difference in their knowledge."
Why models are not like human agents
Krishnan's broader point is that the multi-agent future has been imagined as analogous to human organizations — managers, workers, review, and revision, which map onto hub-spoke — but that AI agents are not like human agents. Models, he argues, are not just models: they carry memories, tools, scaffolds, and execution traces, and choosing which model, scaffold, memory, and tool stack to use is non-trivial. For people, markets work because of local, private knowledge that price signals can elicit, with each person holding private information. For models, "private" knowledge instead emerges from the recursive divergence of prompts, context windows, memories, and tool calls, and the longer they run, the more they specialize. He notes that "it doesn't matter how many memory markdown files you have written, unless you read the right ones at the right time the model behaviour doesn't change."
Krishnan concludes: "Markets will become a true necessity once we hit continual learning but even before that we see models specialise. For now it's more constrained, and there is already a distinct difference. Coase needs Hayek here."
Key claims
- Cost: the market averaged $1.34 per task versus solo's $1.69 and hub-spoke's $5.33 (Source: this essay).
- Quality: the market and solo both averaged 7.2/10, against hub-spoke's 6.7/10 (Source: this essay).
- Coding routing: the market routed 9 of 15 coding runs to GPT-5.2, and 4 runs never received a bid (Source: this essay).
- Reasoning mechanism: brittle problems reward independent retry, with the correct 10/33 answer appearing only in market runs — corroborated by the MarketBench paper.
Provenance and caveats
The essay is a position with empirical support rather than a peer-reviewed paper. Krishnan flags the small sample of 15 tasks himself, and notes that the market's win comes despite poor agent self-assessment rather than because bidding has been solved. The code is published at github.com/strangeloopcanon/hub-vs-spoke.
Relationships
- depends-on: MarketBench: Evaluating AI Agents as Market Participants — Fradkin & Krishnan (April 2026) — the empirical paper this essay extends
- depends-on: The Coasean Singularity? Demand, Supply, and Market Design with AI Agents — Shahidi, Rusak, Manning, Fradkin, Horton (2025) — formal NBER chapter on AI agents as market participants
- supports: Agent Architecture Patterns — empirical evidence that hub-spoke isn't always optimal
- related: Principal-Agent Problem Applied to AI
- related: AI Coding Agents — coding-as-stateful-task explanation for why solo beats hub-spoke and market on most coding work